Making Real-Time Brain Tumor Diagnoses a Reality: AI System Classifies Gliomas From Cryosection Images

by Christos Evangelou, MSc, PhD – Medical Writer and Editor

In a recent study, researchers at Harvard Medical School, Perelman School of Medicine at the University of Pennsylvania, and Cedars-Sinai Medical Center developed a machine learning method for cryosection pathology. The proposed model allowed them to accurately classify gliomas using digital images of cryosection samples during surgery, demonstrating the potential of AI for enabling real-time, intraoperative brain cancer diagnosis.1

“The clinical challenge of making accurate brain cancer during surgery drove us to initiate this study,” said lead researcher Kun-Hsing “Kun” Yu, MD, PhD, assistant professor at Harvard Medical School.

He added that the proposed AI models provide real-time brain cancer diagnoses and molecular subtype prediction, which will assist neurosurgeons in making treatment decisions.

The report was published in the journal Med.

Study Rationale: Overcoming Cryosection Imaging Limitations

Surgeons removing brain tumors rely on quick and accurate diagnosis of tissue samples during surgery to guide decisions on how extensively to remove cancerous growths. Quick-freezing tissue samples for sectioning, known as cryosectioning, allows preliminary pathology evaluation to inform clinical decision-making during surgery.

Explaining what this process is like, Dr. Yu said, “Neurosurgeons often send pieces of brain cancer samples to pathologists during surgery in order to determine the extent of surgical resection and the benefit of additional treatments. Pathologists then process the tissue samples using a special laboratory technique, cryosection, and visually evaluate the processed tissue under the microscope within 10–15 minutes.”

He added that although cryosectioning is faster than standard histology methods, such as H&E staining of formalin-fixed tissues, the quality of the cryosection samples is much lower. Traditional cryosectioning methods often produce unclear images with ice crystal artifacts and tissue irregularities that neuropathologists sometimes struggle to interpret during surgery, potentially leading to misdiagnoses and suboptimal clinical outcomes.

In addition, incorporating molecular profiling per the latest WHO guidelines on brain tumor classification requires genetic tests that take days to weeks. With the WHO Classification of Brain Tumors now requiring genetic data, such as IDH mutation status, cryosection histology alone cannot determine official diagnostic categories.

Approach: Integrating Computer Vision and Deep Learning on Whole-Slide Cryosection Images

To overcome the limitations of current cryosection pathology methods, researchers have developed a model that integrates computer vision and deep learning on whole-slide cryosection images. They termed this model ‘Cryosection Histopathology Assessment and Review Machine’ — or simply ‘CHARM’.

CHARM is a context-aware machine learning method that allows glioma diagnosis during surgery by employing a hierarchical vision transformer as the backbone of the machine learning algorithm. To minimize the manual efforts required for annotating the digital pathology slides, the researchers used a weakly supervised learning approach to analyze the whole-slide images and assign slide-level labels, such as histologic grade or genomic biomarkers, to all tiles sampled from a given slide.

During model training, class-balanced batch sampling and class-balanced loss function were used to minimize instance imbalance of the datasets. Subsequently, the model attributed a separate prediction to each tile, and the median value of all tiles was used as the prediction for the patient.

This aggregation approach in training and prediction phases requires less parameterization, allowing the model to make robust predictions of molecular diagnoses of glioma based on pathology imaging patterns in cryosection slides of brain tissues.

“Simply put, our models learned to diagnose brain cancers using thousands of brain cancer samples and their final diagnoses, which were consensus diagnoses by a panel of experienced pathologists using higher-quality microscopic images that required a few days to process, as well as genetic testing results,” Dr. Yu explained.

He added that by integrating information from both the key diagnostic signals from cancer cells in the images and the surrounding tissues (i.e., contexts) in its decision-making process, they showed that CHARM attained better predictive performance than standard AI methods.

“This approach makes our AI system much more robust than standard AI when testing on samples with lower quality and those collected from different hospitals,” he noted.

Prediction of WHO Subtypes

The team evaluated the performance of CHARM using samples from over 1,500 patients with glioma across three independent cohorts. The proposed AI system achieved near-perfect detection of cancerous tissue, outperforming state-of-the-art convolutional neural networks (CNNs). CHARM provided an area under the curve (AUC) value of 0.98 in identifying malignant cells, compared with an AUC of 0.88 for conventional CNNs.

Glioma surgery aims to maximize tumor removal without damaging healthy brain regions. More complete resection improves survival for certain molecular subtypes but less so for others. Thus, achieving accurate WHO classification during surgery could better inform surgical goals.

Using CHARM, the team was able to accurately classify histologic grades and the three main glioma subtypes per the 2021 WHO guidelines: IDH-mutant astrocytoma, IDH-mutant/1p19q-codeleted oligodendroglioma, and IDH-wildtype glioblastoma.

“Our method took less than a second once the samples were collected and digitized, while the standard molecular profiling methods usually take a few days,” Dr. Yu said. He added that CHARM can enhance diagnostic accuracy by integrating information from the environment surrounding the cancer cells.

The authors further concluded that CHARM has significant potential for providing augmented cryosection evaluation when expert neuropathologists are unavailable. The AI system could also expedite clinical trial enrollment and diagnosis updating per future WHO classification revisions.

Prediction of Genomic Markers

CHARM reliably predicted key genetic alterations, including mutations in IDH, ATRX, and TP53, based on imaging patterns alone. For instance, CHARM achieved an AUC of 0.79–0.82 for distinguishing IDH-mutant tumors from wild-type tumors, compared to an AUC of 0.60–0.65 for existing CNNs.

These findings suggest that CHARM can potentially improve the accuracy and efficiency of glioma diagnosis during surgery by revealing subtle morphological features associated with the molecular profiles of tumors.

Moving Forward

A current limitation of the proposed model is that slides still require standard tissue processing and preparation of digital images of cryosections. Non-invasive tissue analysis methods, such as intraoperative mass spectrometry, along with emerging rapid molecular diagnosis techniques, could complement and enhance CHARM’s imaging predictions while minimizing time-consuming tissue preparation and digitizing tasks.

The authors noted that extending their approach to other cancers to support intraoperative decision-making for cancer diagnosis and treatment remains an active area of research.

“We are currently expanding our analyses to rare types of brain cancers. We are also initiating collaborations with global partners to identify the best way of deploying medical AI systems to resource-limited settings,” Dr. Yu said.

To foster progress in the field, the authors have made their models and codes freely available for research use in all healthcare centers worldwide and have provided the necessary links in their study report.



  1. Nasrallah MP, Zhao J, Tsai CC, et al. Machine learning for cryosection pathology predicts the 2021 WHO classification of glioma. Med (New York, NY). 2023;4(8):526-540.e4. doi:10.1016/j.medj.2023.06.002

Share This Post

Leave a Reply